Objectives
This report is about noise reduction in spatial domain. One of the objectives of this lab is to recognize different types of noise that can effect images which could be due to different applications. The other purpose and the most important is to find the most suitable filter to be applied on such noisy images in order to suppress noise and recover the original image.
Introduction Image Degradation could occur as a result of different reasons such as motion blur, noise, and camera mis-focus. Noise in digital imagery could be defined as a random variation of intensity levels or colour information in images, and is usually occurs as a result of inherent noise found in electronic circuits. It can be generated by the sensor and circuitry of a scanner or digital camera [1]. Noise may also results from transmission interferences across different media.
Noises
Although there exist different types of noises, the Gaussian and Salt and Pepper noises among others are the ones that we are interested in at this experiment. The primary sources of Gaussian noise in images are due to image acquisition process which involves noise from image sensors due to their temperature and also due to the noise from electronics that is controlling the acquisition [1]. While the main sources of salt and pepper noise is the analog to digital converter that has quantization noise [1]. Also, this sort of noise comes from random bit flips (bit errors) during signal transmission.
Noise Reduction
The purpose of image restoration through filtration is to compensate for the defects which degrade the quality of a digital image. There exist various algorithms and methods of combating image noise and the best performing methodology varies from application to application. In choosing a noise reduction algorithm, available computing power, acceptable levels of image degradation and noise type are common considerations.
Among many filtration methods, mean filters including arithmetic and geometric mean filters are well known techniques that can be used in noise reduction. An arithmetic mean filter gets rid of noise such as Gaussian noise from the image. However, this is achievable through blurring the image by calculating the average of all pixels within a portion of the image specified by the window size [2]. The geometric mean is similar in function to the Arithmetic mean, the difference is that it calculates the product of neighbouring values instead of summation as the Arithmetic does [3]. Wavelet-transform is another method of filtration which de-noise image by selecting a few coefficients containing high image detail while ignoring other coefficients, so due to the fact that the noise is distributed uniformly across all the coefficients, selecting a subset of coefficients with high image detail reduces the noise [4]. Statistical methods for image noise reduction exist as well, though they are not frequently used because of computational complexity.
Methodology
In this lab task, investigation on the images provided (noisy-1, noisy-2, noisy-3 and the original image) will be carried on in order to recognize the type of noise they got affected by. This could be achieved through introducing different noises into the original image until similar results are attained. Next, Arithmetic-mean and geometric-mean filters with different masks will be used to counter the noise affects pertaining to random intensity variations because these filters tend to smooth out intensity variations through averaging.
Median filter is also explored in this task as this filter type is good at removing salt and pepper noise from the images and also cause less blurring at the edge of the image.
The implementation details to recognize the type of noise appliedand and the algorithms used to denoise
the given images is as follows:
Discussion
After applying different noises on the (img05) image with different variables setups, it came to conclusion that image Noisy-1 as shown in (figure-1) was a result of using Gaussian noise with mean= 0.15 on the original image. In (figure-2), we can see that after applying Speckle noise with variance= 0.15 we got the same result in Noisy-2. (Figure-3) shows that Noisy-3 resulted from applying Salt & Pepper noise of percentage equals to 15 on (img05).
Noisy-1 image is a less noisy image relative to the other two imagesas there are less pixel intenisty variarions as compared to the originalimage. Intensity normalization can reduce this type of noise affect, therefore we can use averaging(mean) filters that we have designed to reduce the noise in this image. The results can be seen in (figure-4).
In (figure-5), we can see that Noisy-2 has more variation in pixel intensity values as compared to the
original image and previously seen Noisy-1 image .Also, the pixel intensity variation can be seen in group of contiguous pixels forming spot, hence the name speckle noise [5]. This noise type can also be reduced by intensity normalization on local group of pixels, so applying average filter can reduce image noise.
As the type of noise in Noisy-3 image is salt-and-pepper noise. We can easily suppress this sort of noise using median filtering. Therefore, median filter developed earlier is applied on the image and salt and pepper noise reduction is clear from the displayed results in (figure-6).
For comparison, arithmetic mean filtering and geometric mean filtering is alos applied to this image and both of the averaging filters peformed poorly on salt and pepper noise because arithemtic mean filter averages out the high(255) and low(0) values and putting mid-tones(128) values, whereas, the geometric mean filter has darkened the image because it has a characteristic of multipying pixel values, and multiplying by low pixel values (0 or pepper) makes the image tend to darken.
Conclusion
In this lab task, different noise reduction filters are used to undo the affect of various noise types in the given set of images. For speckle noise, averaging with Arithmetic mean filter resulted into the best output whereas geometric mean filter darkened the overall tone of the image. For salt and pepper type of noise, median filter successfully substituted the extreme pixel intensity values in the corrupted image with midtone values. This shows that understanding the type of noise and selecting an appropriate filtering technique based on that determines the success of image filtering.
Sample Solution